Abstract

Rapid yet accurate post-earthquake damage assessment of highway bridges is essential to ensure transportation network resilience. Traditional inspection, fragility analysis, and finite element modeling are time-consuming or inaccurate for real-time decision-making. This study compares ten machine-learning classifiers for the seismic damage states of highway steel-girder bridges that incorporate directional intensity measures to represent the variance of ground motions. Based on the nonlinear time-history analyses of two Missouri bridges, one Michigan bridge, and one Wisconsin bridge, multi-parameter structural damage indices were proposed and mapped to five damage states from none to collapse. Among ten supervised algorithms studied, the artificial neural network trained using the Missouri bridge data set was most generalizable on the unseen test set from the same bridges, with a prediction accuracy of 0.95 and an area-under-the-curve (AUC) of 0.98. When applied to two unseen bridges from Michigan and Wisconsin, the neural network's average accuracy decreased to approximately 0.81. The directional intensity measures and the multi-parameter damage indices enabled robust and scalable classifications of highway bridges, even with domain shifting, and thus the rapid post-earthquake damage assessment of a regional transportation network.

Department(s)

Civil, Architectural and Environmental Engineering

Publication Status

Full Text Access

Keywords and Phrases

Highway-bridges modeling; Machine learning algorithms; Seismic damage classification; Soil-structure interaction

International Standard Serial Number (ISSN)

1873-7323; 0141-0296

Document Type

Article - Journal

Document Version

Citation

File Type

text

Language(s)

English

Rights

© 2026 Elsevier, All rights reserved.

Publication Date

01 Aug 2026

Download

Full Text Link

Share

 
COinS